A method for early leakage prediction of boiler tubes based on regional noise increment feature identification
By dividing the region into grids and identifying the noise increment characteristics of the acoustic sensor inside the boiler furnace, constructing the noise increment direction vector and spatial gradient analysis, accurate prediction of early leakage is achieved, solving the problems of high false alarm rate and high false alarm rate in the existing technology, and improving the reliability of boiler tube leakage detection.
Patent Information
- Application Number
- CN202611579060.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-13
- Publication Date
- 2026-11-10
AI Technical Summary
Existing boiler tube leakage detection technologies are difficult to accurately identify leaks in the early stages. Single-point sensor detection methods cannot utilize regional spatial features, and fixed thresholds are easily affected by boiler load changes and background noise, leading to misjudgments and missed detections.
By dividing multiple acoustic sensors into regional grids, a frequency domain noise spectrum is generated, a noise increment direction vector is constructed, and the spatial gradient and temporal correlation of the increment direction field matrix are used for tracking to determine the early leakage risk area.
It improves the accuracy of early leak identification, reduces the false alarm rate, can stably output the leak risk area under different operating conditions, and enhances the ability to distinguish between low-energy leak signals and random noise disturbances.
Smart Images

Figure CN123065682A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of boiler tube leakage monitoring technology, specifically a method for early leakage prediction of boiler tubes based on regional noise increment feature identification. Background Technology
[0002] In boiler tube leak detection, acoustic sensors are commonly used to receive noise signals from inside the furnace. Leaks are identified by monitoring changes in sound pressure level, spectral energy, or power in specific frequency bands. The combustion process inside the furnace itself generates broadband, strong noise, and interference sources such as flue gas flow, sootblower operation, and fuel input can also cause fluctuations in the noise spectrum. In the early stages of leakage, the acoustic emission energy generated by microcracks or pinholes in the furnace tubes is weak, and the spectral components easily overlap with the background combustion noise, resulting in significant random fluctuations in the amplitude and energy characteristics obtained by a single sensor. Existing technologies mainly rely on fixed thresholds or historical statistical thresholds to determine the time-domain amplitude and frequency-domain energy of the acoustic signal. When boiler load changes, air distribution is adjusted, or coal type changes, the overall spectrum of background noise shifts, making it difficult for fixed thresholds to adapt to dynamic conditions and easily misjudging normal combustion disturbances as leaks. Detection methods based on single points or a small number of sensors do not form regional spatial characteristics and cannot utilize boundary differences in noise changes between different monitoring areas to locate the abnormal range, limiting their ability to distinguish between localized early leaks and global combustion state fluctuations. Some methods spatially interpolate or average the spectral amplitudes of multiple acoustic sensors. The amplitude variation after regional averaging smooths the local directional characteristics, further weakening the subtle differences in early leakage at local boundaries. Frequency domain difference or trend analysis typically only focuses on amplitude magnitude, failing to express the spatial arrangement of frequency increase / decrease directions or track the symbol stability of these directions over continuous sampling. The noise increment direction caused by early leakage often forms a identifiable spatial discontinuity at local boundaries, while the direction of random noise disturbances exhibits frequent flips. Existing technologies lack the extraction of noise increment direction gradients at regional boundaries and the utilization of the temporal flipping characteristics of direction vectors, thus limiting the predictive ability for early, weak leakage. Summary of the Invention
[0003] The purpose of this invention is to provide a method for predicting early leakage of boiler tubes based on regional noise increment feature identification. The method involves dividing multiple acoustic sensors into regional grids and generating frequency domain noise spectra. Noise increment direction vectors are constructed by differential symbols at each frequency point. Candidate abnormal regions are extracted using the spatial gradient of the increment direction field matrix. Then, the noise increment direction vectors at continuous sampling times are time-series correlated and tracked. Early leakage risk regions are determined based on the symbol flipping frequency, thereby reducing the interference of boiler background noise and operating condition fluctuations on the prediction of early leakage of boiler tubes.
[0004] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a method for predicting early leakage of boiler tubes based on regional noise increment feature identification, comprising:
[0005] Multiple acoustic sensors deployed inside the boiler furnace are divided into regional grids according to spatial coordinates to obtain sensor arrays for each monitoring area;
[0006] The time-domain noise signals of the acoustic sensors in each monitoring area are collected synchronously, and a fast Fourier transform is performed on each time-domain noise signal to obtain the frequency-domain noise spectrum of each monitoring area.
[0007] Within each monitoring area, the current frequency domain noise spectrum is compared with the frequency domain noise spectrum at the previous sampling time by frequency point differential calculation to obtain the differential amplitude at each frequency point, and the positive and negative sign sequence of the differential amplitude at each frequency point is counted to generate the noise increment direction vector of the area.
[0008] An incremental direction field matrix is constructed using the noise increment direction vectors of all monitoring areas. Spatial gradient operation is performed on the incremental direction field matrix to obtain the connected domain of the region boundary where the direction gradient magnitude exceeds the pre-judgment threshold. The monitoring area covered by this connected domain is marked as a candidate anomaly region.
[0009] Within the candidate anomaly region, the noise increment direction vectors at multiple consecutive sampling times are time-series correlated and tracked. When the sign flip frequency of the tracked noise increment direction vector is lower than the dynamic threshold, the candidate anomaly region is determined to be an early leakage risk region.
[0010] Furthermore, the multiple acoustic sensors deployed inside the boiler furnace are divided into regional grids according to spatial coordinates to obtain sensor arrays for each monitoring area, specifically including:
[0011] Extract the three-dimensional spatial installation coordinates of all acoustic sensors in the boiler furnace, and divide the acoustic sensors into several sensor groups with the constraint that the Euclidean distance between adjacent sensors does not exceed a preset spacing threshold.
[0012] Take the smallest circumscribed cuboid of the three-dimensional spatial installation coordinates of all acoustic sensors in each sensor group, divide the smallest circumscribed cuboid into equal-distance grids in the horizontal and vertical directions to obtain several grid cells, and merge the grid cells containing at least one sensor into the same monitoring area.
[0013] The sensor numbers of the acoustic sensors contained in each monitoring area are associated and stored to form a sensor array index table corresponding to each monitoring area.
[0014] Furthermore, the frequency domain noise spectrum of each monitoring area is obtained, specifically including:
[0015] According to the sensor array index table, the analog-to-digital converters of all acoustic sensors in each monitoring area are synchronously triggered with the same sampling clock source, and the discrete time-domain sampling sequence within a preset time period is collected as the time-domain noise signal of the sensor.
[0016] The time-domain noise signals of each acoustic sensor are truncated by applying a Hanning window function. The truncated time-domain sequences are then padded with zeros to the preset number of Fourier transform points and a fast Fourier transform is performed to obtain the initial frequency domain sequences of each sensor.
[0017] The frequency domain amplitude of the monitoring area at the same frequency point is obtained by taking the arithmetic mean of the amplitudes of the initial frequency domain sequences of all sensors in the same monitoring area at the same frequency point. After aggregating all frequency points, the frequency domain noise spectrum of each monitoring area is formed.
[0018] Furthermore, the noise increment direction vector for this region is generated, specifically including:
[0019] For the same monitoring area, extract the frequency domain noise spectrum vector at the current sampling time and the frequency domain noise spectrum vector at the previous sampling time, subtract the frequency domain amplitude at the same frequency index in the two vectors, and obtain the differential amplitude corresponding to each frequency index.
[0020] For each frequency index, a sign function is used to calculate the differential amplitude. When the differential amplitude is greater than zero, the sign function outputs a positive sign; when the differential amplitude is less than zero, the sign function outputs a negative sign; and when the differential amplitude is equal to zero, the sign function outputs zero. The sign function outputs of all frequency indices are arranged in ascending order of frequency index to form the noise increment direction vector of the monitoring area.
[0021] Furthermore, an incremental direction field matrix is constructed using the noise increment direction vectors of the entire monitoring area, specifically including:
[0022] The noise increment direction vectors of each monitoring area are used as matrix elements and arranged into a two-dimensional increment direction field matrix according to the spatial adjacency topology of each monitoring area. The row and column indices of this matrix correspond to the sequence numbers of the monitoring areas in the horizontal and vertical directions, respectively.
[0023] Furthermore, spatial gradient operations are performed on the incremental direction field matrix to obtain the connected regions whose directional gradient magnitudes exceed a predetermined threshold, specifically including:
[0024] For each element in the two-dimensional incremental direction field matrix, the Hausdorff distance between its direction vector and the direction vector of its right-hand neighbor and its lower neighbor is taken as the horizontal gradient and vertical gradient, respectively, and the gradient magnitude of the matrix element is synthesized from the horizontal gradient and the vertical gradient.
[0025] The monitoring areas corresponding to matrix elements whose gradient magnitude is greater than the predicted threshold are extracted. The four-connected neighborhood growth algorithm is used to merge adjacent extracted monitoring areas into connected regions. All monitoring areas covered by the connected region with the smallest area are marked as candidate abnormal regions.
[0026] Furthermore, a four-connected neighborhood growth algorithm is used to merge adjacent extracted monitoring regions into connected components. The entire monitoring region covered by the connected component with the smallest area is marked as a candidate anomaly region, specifically including:
[0027] Initialize a labeling matrix with the same size as the two-dimensional incremental direction field matrix. The initial value of each element in the labeling matrix is unlabeled. Traverse the two-dimensional incremental direction field matrix in row-first and column-second order. When the gradient magnitude is greater than the prediction threshold and the corresponding position in the labeling matrix is unlabeled, create a new connected component number.
[0028] Using the matrix element as the seed point for growth, check the matrix elements adjacent to the seed point in the four directions of up, down, left, and right. When the gradient magnitude of the adjacent matrix element is also greater than the pre-judgment threshold, the adjacent matrix element is assigned to the current connected component number and the corresponding position in the label matrix is updated to the labeled state. Then, the newly assigned matrix element is used as the new seed point to repeat the growth process.
[0029] Count the number of matrix elements contained under each connected component number, take all the monitoring areas corresponding to the connected component with the fewest matrix elements as the marking objects, and mark these monitoring areas as candidate abnormal areas.
[0030] Furthermore, within the candidate anomaly region, temporal correlation tracking is performed on the noise increment direction vectors at multiple consecutive sampling times, specifically including:
[0031] Using the noise increment direction vector of the candidate anomaly region at the current sampling time as a reference, search for the noise increment direction vector with the smallest angle with the reference direction vector in the candidate anomaly region set at the previous sampling time. If the smallest angle with the reference direction vector is less than a preset angle threshold, then the two direction vectors are associated as consecutive nodes on the same tracking chain.
[0032] Increment the duration of the successfully associated tracking chain by one, and count the symbol sequence of the noise increment direction vector at each sampling moment within the current duration of the tracking chain. Calculate the ratio of the number of times adjacent symbols in the symbol sequence flips to the total length of the symbol sequence to obtain the current symbol flip frequency of the tracking chain.
[0033] Furthermore, the current symbol flip frequency of the tracing chain is obtained, specifically including:
[0034] For each sampling moment corresponding to the noise increment direction vector stored on the tracking chain, extract the frequency index positions where the symbol function output is non-zero in each vector, take the symbol function output values at these frequency index positions and arrange them in ascending order of frequency index to form the symbol subsequence of that sampling moment, and splice the symbol subsequences of multiple consecutive sampling moments on the tracking chain in chronological order to form the overall symbol sequence of the tracking chain;
[0035] Traverse the symbol values of adjacent positions in the overall symbol sequence. Whenever an adjacent symbol value changes from positive to negative or vice versa, increment the flip count by one. After traversing the sequence, divide the flip count by the length of the overall symbol sequence and decrement by one to obtain the current symbol flip frequency of the tracing chain.
[0036] The current symbol flip frequency is temporarily stored in the status register of the tracking chain, serving as a historical reference value for comparison of the symbol flip frequency at the next sampling moment.
[0037] Furthermore, when the sign-flipping frequency of the tracked noise increment direction vector is lower than the dynamic threshold, the candidate anomaly region is determined to be an early leakage risk region, specifically including:
[0038] The current symbol flip frequency is compared with the preset lower limit of the flip frequency. When the current symbol flip frequency is lower than the preset lower limit of the flip frequency, the candidate abnormal area is identified as an early leakage risk area.
[0039] Furthermore, the current symbol flip frequency is compared with a preset lower limit. When the current symbol flip frequency is lower than the preset lower limit, the candidate anomaly region is identified as an early leakage risk region, specifically including:
[0040] Calculate the arithmetic mean of the symbol flip frequencies of all tracking chains within the candidate anomaly region at the current sampling time, and use this arithmetic mean as the regional average flip frequency of the candidate anomaly region.
[0041] Using the average flip frequency of the candidate anomaly region at the previous sampling time as the benchmark value, the average flip frequency of the region at the current sampling time is divided by the benchmark value to obtain the flip frequency attenuation ratio of the candidate anomaly region.
[0042] When the attenuation ratio of the flip frequency is less than the preset attenuation threshold, the region identifier of the candidate abnormal region and the corresponding sensor array index table are output as the determination result of the early leakage risk region.
[0043] Furthermore, the prediction threshold is dynamically determined by the weighted sum of the historical mean and standard deviation of the gradient amplitude of each monitoring area under normal boiler operating conditions.
[0044] Furthermore, the Hausdorff distance is used to measure the maximum degree of difference between two direction vectors.
[0045] Furthermore, in the four-connected neighborhood growth algorithm, growth in that direction is stopped when a boundary or a marked element is encountered during the growth process.
[0046] Furthermore, the preset angle threshold ranges from 0° to 90°, which is used to limit the upper limit of the allowable deviation of the angle between direction vectors in time-series correlation tracking.
[0047] The technical effects and advantages provided by the present invention in the above technical solution are as follows:
[0048] The frequency-by-frequency differential sign of the frequency domain noise spectrum is used to construct the noise increment direction vector, shifting the description of regional noise changes from amplitude comparison to spatial encoding of spectral increase / decrease directions. The sign of the differential amplitude depends only on the relative change direction of spectral energy at the same frequency point at adjacent sampling times, and is insensitive to amplitude scaling caused by differences in acoustic sensor gain, overall shift of the furnace background spectrum, and load conditions. After arranging the noise increment direction vectors of all monitored areas into an increment direction field matrix, the Hausdorff distance between the noise increment direction vectors of adjacent areas serves as the spatial gradient, which can characterize the degree of discontinuity in noise change patterns at the regional boundaries. Matrix elements with directional gradient amplitudes exceeding a pre-defined threshold are grown into connected domains at the regional boundaries through four-connected neighborhood growth. The monitored areas covered by these connected domains are marked as candidate anomaly regions. Through spatial gradient and connected domain growth, the marking of candidate anomaly regions no longer depends on a pre-set leakage amplitude threshold, but is determined by the degree of discontinuity in the noise increment direction at the regional boundaries. This distinguishes the change in noise increment direction caused by local early leakage from fluctuations in the overall boiler combustion state, avoiding false alarms and false negatives caused by background noise interference from a single sensor amplitude threshold. Within the candidate anomaly region, the direction vectors of noise increments at multiple consecutive sampling times are temporally correlated and tracked. A tracking chain is established based on the minimum angle between the direction vectors of adjacent sampling times as the correlation condition. The symbol flip frequency is obtained by statistically analyzing the ratio of the number of times the positive and negative signs flip in the symbol sequence to the total length of the symbol sequence. Local acoustic emissions caused by early leakage are characterized in the frequency domain by continuous increases and decreases in the same direction at several frequency points, with a low and stable trend in the direction vector symbol flip frequency. Spectral fluctuations caused by random combustion disturbances and instantaneous airflow impacts are characterized by frequent reversals in the direction of increase and decrease at each frequency point, with a higher symbol flip frequency. The current symbol flip frequency is compared with a dynamic threshold, and the early leakage risk region is output when the attenuation ratio of the regional average flip frequency meets the condition. The dynamic threshold adapts to the temporal fluctuation range of background noise in different operating periods. The stability information of the temporal direction is incorporated into the judgment process, enhancing the ability to distinguish between low-energy continuous leakage signals and random noise disturbances, and enabling the detection of early leakage risks without relying on an increase in absolute amplitude. Attached Figure Description
[0049] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.
[0050] Figure 1 This is a flowchart of a method for predicting early leakage of boiler tubes based on regional noise increment feature identification;
[0051] Figure 2 This is a flowchart of the noise increment direction vector generation process. Detailed Implementation
[0052] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0053] See Figure 1 This invention provides a method for predicting early leakage of boiler tubes based on regional noise increment feature identification, comprising: dividing multiple acoustic sensors deployed in the boiler furnace into regional grids according to spatial coordinates to obtain sensor arrays for each monitoring area; synchronously acquiring time-domain noise signals from acoustic sensors in each monitoring area, performing a fast Fourier transform on each time-domain noise signal to obtain the frequency-domain noise spectrum of each monitoring area; within each monitoring area, performing frequency-point differential calculation between the current frequency-domain noise spectrum and the frequency-domain noise spectrum at the previous sampling time to obtain the differential amplitude at each frequency point, and statistically analyzing the positive and negative sign sequences of the differential amplitude at each frequency point to generate a noise increment direction vector for that area; constructing an increment direction field matrix using the noise increment direction vectors of all monitoring areas, performing spatial gradient operation on the increment direction field matrix to obtain the connected domain of the region boundary where the direction gradient amplitude exceeds the prediction threshold, and marking the monitoring area covered by the connected domain as a candidate abnormal region; within the candidate abnormal region, performing temporal correlation tracking of the noise increment direction vectors at multiple consecutive sampling times, and determining the candidate abnormal region as an early leakage risk region when the sign flip frequency of the tracked noise increment direction vector is lower than a dynamic threshold.
[0054] In the overall implementation process, acoustic sensors within the boiler furnace are divided into different monitoring areas using spatial coordinates, with each monitoring area corresponding to a sensor array. After the sensor arrays synchronously acquire time-domain noise signals, the frequency-domain noise spectrum of each monitoring area is obtained by averaging the frequency-domain sequences of the corresponding sensor arrays. The frequency-by-frequency difference between the current frequency-domain noise spectrum and the frequency-domain noise spectrum at the previous sampling time generates a noise increment direction vector. The noise increment direction vectors of each monitoring area are arranged according to spatially adjacent topological relationships to form an increment direction field matrix. After performing spatial gradient calculations on this matrix, the connected domains of the region boundaries are extracted. Within candidate abnormal regions, the symbol flip frequency is calculated based on the temporal correlation tracking results of the noise increment direction vectors at adjacent sampling times. Candidate abnormal regions with symbol flip frequencies lower than the dynamic threshold are identified as early leakage risk regions. A comparative experiment was conducted to verify the technical effect of the combination of the above-mentioned techniques. In the experiment, the traditional single-sensor sound pressure level threshold detection method, the detection method based on regional average spectral energy, and the method of this invention were used to identify the early leakage conditions of the same boiler tube. The results show that when changes in boiler load, air distribution, or coal type cause an overall shift in the background noise spectrum, traditional single-sensor sound pressure level threshold detection methods are prone to numerous false alarms, while the regional average spectral energy method struggles to identify localized weak leaks due to averaging and smoothing effects. Our proposed method eliminates random fluctuations from single sensors through regional gridding and synchronous acquisition, highlights the directional changes in the spectrum caused by leaks by constructing a noise increment direction vector through frequency-point differential symbolization, separates localized abnormal regions from global background fluctuations through spatial gradient connected component extraction, and further eliminates random transient interference through temporal correlation tracking and symbol flipping frequency determination. Compared to traditional methods, this method significantly improves the detection capability for early-stage weak leaks while maintaining a low false alarm rate. It can identify minute pinhole leak signals that traditional methods cannot detect, and it can stably output leak risk areas under different boiler load conditions.
[0055] In practical implementation, the multiple acoustic sensors deployed inside the boiler furnace are divided into regional grids according to spatial coordinates, and the sensor arrays for each monitoring area are obtained as follows:
[0056] The system reads the installation records of all acoustic sensors inside the boiler furnace. These records include the sensor number and three-dimensional spatial coordinates of each sensor. The three-dimensional spatial coordinates are based on a preset reference point inside the boiler furnace and consist of horizontal, vertical, and depth coordinates. A one-to-one correspondence is established between the sensor number and the three-dimensional spatial coordinates of each acoustic sensor, and this correspondence is stored in the data processing unit.
[0057] Each acoustic sensor is treated as a node in a graph structure, and the Euclidean distance between any two acoustic sensors is calculated. When the Euclidean distance between any two acoustic sensors is less than or equal to a preset spacing threshold, an edge is established between the nodes corresponding to the two acoustic sensors. All connected components in the graph structure are extracted, and the acoustic sensors corresponding to the nodes in each connected component constitute a sensor group. Within a sensor group, the Euclidean distance between any acoustic sensor and at least one other acoustic sensor within the group does not exceed the preset spacing threshold, and the Euclidean distance between any two acoustic sensors in different sensor groups is greater than the preset spacing threshold. The preset spacing threshold is stored in the data processing unit and is pre-set based on the installation density of the acoustic sensor array in the boiler furnace and the acoustic wave propagation attenuation characteristics. The preset spacing threshold is less than or equal to the maximum allowable spatial interval between adjacent acoustic sensors in the boiler furnace.
[0058] For each sensor group, determine the maximum and minimum horizontal coordinates, the maximum and minimum vertical coordinates, and the maximum and minimum depth coordinates of all acoustic sensors in the group. Six boundary surfaces are then defined by these values. The cuboid formed by these six boundary surfaces constitutes the minimum circumscribed cuboid of the sensor group.
[0059] The minimum circumscribed cuboid of each sensor group is divided into equidistant grids in both the horizontal and vertical directions. The minimum circumscribed cuboid is divided along its horizontal length by a preset horizontal spacing and along its vertical length by a preset vertical spacing. No division is performed in the depth direction, resulting in several grid cells. Both the preset horizontal and preset vertical spacing are set to the same value as a preset spacing threshold.
[0060] Determine whether each grid cell contains the three-dimensional spatial coordinates of at least one acoustic sensor. Grid cells containing at least one acoustic sensor are marked as valid grid cells. Valid grid cells that are adjacent in the horizontal or vertical directions are merged into the same monitoring area. All valid grid cells within the same monitoring area are interconnected.
[0061] A monitoring area number is established, and the sensor numbers of all acoustic sensors falling within the same monitoring area are associated with that monitoring area number, generating a sensor array index table. The sensor array index table records the correspondence between the monitoring area number and the set of sensor numbers. This table is stored in the data processing unit for subsequent retrieval of the sensor number set based on the monitoring area number.
[0062] In practice, the following methods are used to obtain the frequency domain noise spectrum of each monitoring area:
[0063] According to the sensor array index table, the analog-to-digital converters (ADCs) of all acoustic sensors in each monitoring area are synchronously triggered by the same sampling clock source. The same sampling clock source is a global clock generator, which outputs a square wave signal of the same frequency to the trigger inputs of the ADCs of all acoustic sensors in each monitoring area. Synchronous triggering uses the rising edge of the square wave signal. When the rising edge of the same clock arrives, the ADCs of all acoustic sensors in each monitoring area simultaneously begin converting the analog voltage signals output by the acoustic sensors into discrete digital sampled values. The frequency of the square wave signal output by the sampling clock source is the same as the sampling rate of the ADC, which is set to be greater than or equal to twice the highest frequency component of the acoustic signal from early leakage in the boiler tubes. A preset duration is stored in the data processing unit, and the preset duration is greater than or equal to the length of a complete cycle of the lowest frequency component of the acoustic signal from early leakage in the boiler tubes. Each acoustic sensor's ADC generates a fixed number of discrete digital sampled values within the preset duration. These discrete digital sampled values are arranged in chronological order of sampling time, forming the time-domain noise signal of the acoustic sensor.
[0064] A Hanning window function is applied to truncate the time-domain noise signal of each acoustic sensor. The window length of the Hanning window function is the same as the number of discrete digital samples of the time-domain noise signal of the acoustic sensor. The window coefficient of the Hanning window function at the discrete sampling points is calculated using the following formula:
[0065]
[0066] in, This indicates that the Hanning window function is in the first position. The window coefficients at each discrete sampling point are used for weighted truncation of the time-domain signal; For discrete sampling point numbers, The range of values is arrive Integers; Let be the window length of the Hanning window function. It is equal to the number of discrete digital samples of the time-domain noise signal of the acoustic wave sensor; Pi Values ; It is a cosine function; Let be the independent variable of the cosine function, representing the th... The phase angle value corresponding to each discrete sampling point.
[0067] The window coefficient of the Hanning window function is multiplied point-by-point with the corresponding discrete digital sample values in the time-domain noise signal of the acoustic sensor to obtain the truncated time-domain sequence. The truncated time-domain sequence is then padded with zeros to a preset number of Fourier transform points. This preset number of Fourier transform points is stored in the data processing unit and is greater than the number of discrete digital sample values in the time-domain noise signal of the acoustic sensor. The preset number of Fourier transform points is set to... The integer power of 0. The zero-padding method involves appending zero-value discrete sampling points to the end of the truncated time-domain sequence, making the length of the zero-padding sequence equal to the preset number of Fourier transform points.
[0068] Perform a Fast Fourier Transform (FFT) on the zero-padded sequence. The FFT uses a base-based approach. The time-decimation algorithm takes a zero-padded real number sequence as input to the Fast Fourier Transform (FFT) and outputs a frequency domain sequence in complex form. Each frequency point in the complex frequency domain sequence corresponds to a complex number, and the magnitude of the complex number is used as the amplitude at that frequency point. The initial frequency domain sequence for each sensor contains half the preset number of Fourier Transform points plus one effective frequency point, with the effective frequency point index starting from... Up to half the preset number of Fourier transform points.
[0069] The arithmetic mean of the amplitudes of the initial frequency domain sequences of all sensors within the same monitoring area at the same frequency point is taken. For monitoring areas numbered... The monitoring area, at the frequency index is The monitoring area is numbered as follows: The monitoring area contains the following number of acoustic sensors: The monitoring area number is The first monitoring area The initial frequency domain sequence of each acoustic sensor is indexed at the frequency point. The amplitude at that point is The monitoring area number is The monitoring area in the frequency index frequency domain amplitude Calculate using the following formula:
[0070]
[0071] in, Indicates the monitoring area number; Indicates frequency point index, The range of values is An integer up to half the preset number of Fourier transform points; The monitoring area number is The number of acoustic sensors included in the monitoring area. The value is greater than or equal to Integers; The monitoring area number is The serial number of the acoustic sensor within the monitoring area. The range of values is arrive Integers; The monitoring area number is The first monitoring area The initial frequency domain sequence of each acoustic sensor is indexed at the frequency point. The amplitude at that point; The monitoring area number is The monitoring area in the frequency index The frequency domain amplitude at that point.
[0072] The arithmetic mean of all frequency indexes is calculated one by one to obtain the monitoring area number. The frequency domain amplitude values of the monitoring area at each frequency point are calculated, and the frequency domain amplitude values of all frequency points are arranged in ascending order according to the frequency point index to form a monitoring area numbered as follows: The frequency domain noise spectrum of the monitoring area is obtained. The same operation is performed on each monitoring area number to obtain the frequency domain noise spectrum of each monitoring area.
[0073] In specific implementation, please refer to Figure 2 The implementation method for generating the noise increment direction vector for each monitoring area is as follows:
[0074] The data processing unit stores the frequency domain noise spectrum vectors of each monitoring area at each sampling time. Each frequency domain noise spectrum vector consists of multiple frequency point amplitudes, which are arranged in ascending order according to their frequency point indices. The range of frequency point indices corresponds to the preset number of Fourier transform points. The frequency domain noise spectrum vector of the same monitoring area at the current sampling time has the same set of frequency point indices and the same vector length as the frequency domain noise spectrum vector at the previous sampling time.
[0075] For the same monitoring area, extract the frequency domain noise spectrum vector at the current sampling time and the frequency domain noise spectrum vector at the previous sampling time. Subtract the frequency domain amplitudes at the same frequency index in the frequency domain noise spectrum vectors at the current and previous sampling times to obtain the differential amplitude corresponding to each frequency index. In the monitoring area numbered... In the monitoring area, the frequency index is The difference magnitude at the point is calculated using the following formula:
[0076]
[0077] in, The monitoring area number is The monitoring area is at the frequency index of The differential amplitude at the specified point is expressed in units of frequency domain amplitude. The differential amplitude can be positive, negative, or zero. Indicates frequency point index, The range of values is An integer up to half the preset number of Fourier transform points; Indicates the monitoring area number, The value range is the set of monitoring area numbers recorded in the sensor array index table; Indicates the current sampling time sequence number. greater than or equal to Positive integers; This indicates the sequence number of the previous sampling time, which is the same as the sequence number of the current sampling time. minus The resulting positive integer; The monitoring area number is The monitoring area at the current sampling time number The frequency index of the frequency domain noise spectrum vector is Frequency domain amplitude at that point; The monitoring area number is The monitoring area is at the previous sampling time number The frequency index of the frequency domain noise spectrum vector is The frequency domain amplitude at that point.
[0078] In some embodiments, a sign function is performed on the differential amplitude value corresponding to each frequency index. The input to the sign function is the differential amplitude value corresponding to the frequency index, and the output is the sign function output value. When the differential amplitude value is greater than zero, the sign function outputs a positive sign, indicated by the sign "". "" indicates that when the difference magnitude is less than zero, the sign function outputs a negative sign, which is represented by the sign "". "" indicates that when the difference magnitude is equal to zero, the sign function outputs a zero sign, which is represented by the sign "". "express.
[0079] Arrange the symbol function outputs of all frequency indexes in ascending order of frequency index to form the monitoring area number. The noise increment direction vector for the monitoring area is calculated. The dimension of the noise increment direction vector is the same as the number of frequency points in the frequency domain noise spectrum vector. Each element in the noise increment direction vector corresponds to the symbol function output value at a frequency index position. The above difference calculation and symbol function operation are repeated for each monitoring area to obtain the noise increment direction vector for each monitoring area.
[0080] In some embodiments, the noise increment direction vector of each monitoring area is stored in the data processing unit and indexed by the sampling time sequence number and the monitoring area number. The sign function output value in the noise increment direction vector is stored continuously in ascending order by the frequency point index.
[0081] In practice, the noise increment direction vectors of each monitoring area are used as matrix elements and arranged into a two-dimensional increment direction field matrix according to the spatial adjacency topology of each monitoring area. The row index of the two-dimensional increment direction field matrix corresponds to the vertical sequence number of the monitoring area, and the column index corresponds to the horizontal sequence number of the monitoring area. The spatial adjacency topology of the monitoring areas is determined based on their adjacent positions in the horizontal and vertical directions within the boiler furnace. Horizontally adjacent monitoring areas are located in the same row and adjacent columns in the two-dimensional increment direction field matrix, while vertically adjacent monitoring areas are located in the same column and adjacent rows.
[0082] For each element in the two-dimensional incremental direction field matrix, the Hausdorff distance between the direction vectors of the current element and its right-hand neighbor is taken as the horizontal gradient of the current element. The current element is the element located at the […]. Line number When dealing with matrix elements in a column, the right-hand adjacent element of the current matrix element is the element located at the th position in the two-dimensional incremental direction field matrix. Line number The matrix elements of the column. The Hausdorff distance between the current matrix element and its right-side neighbor is calculated as follows: Extract the frequency index positions whose sign function output value is positive or negative from the direction vector corresponding to the current matrix element to form a first frequency index set; extract the frequency index positions whose sign function output value is positive or negative from the direction vector corresponding to the right-side neighbor element to form a second frequency index set; calculate the minimum Euclidean distance between each frequency index in the first frequency index set and all frequency indices in the second frequency index set, and record the maximum value among the minimum values corresponding to all frequency indices in the first frequency index set as the first one-way Hausdorff distance; calculate the minimum Euclidean distance between each frequency index in the second frequency index set and all frequency indices in the first frequency index set, and record the maximum value among the minimum values corresponding to all frequency indices in the second frequency index set as the second one-way Hausdorff distance; take the larger of the first and second one-way Hausdorff distances as the Hausdorff distance between the current matrix element and its right-side neighbor element. The Euclidean distance between two frequency indexes is the absolute value of the difference between the two frequency indexes.
[0083] Similarly, the Hausdorff distance between the direction vectors of the current matrix element and its next-lower element is taken as the vertical gradient of the current matrix element. The current matrix element is the element located at the [missing element] in the two-dimensional incremental direction field matrix. Line number When dealing with matrix elements in a column, the element immediately below the current matrix element is the element located at the th position in the two-dimensional incremental direction field matrix. Line number The matrix elements of the column. The Hausdorff distance between the current matrix element and its adjacent element below it is calculated in the same way as the Hausdorff distance of the horizontal gradient.
[0084] The gradient magnitude of the current matrix element is synthesized from the horizontal and vertical gradients, and the gradient magnitude is calculated according to the following formula:
[0085]
[0086] in, Indicates the position of the two-dimensional incremental direction field matrix located at the th Line number The gradient magnitude of the matrix elements in the column, with the gradient magnitude in units of Hausdorff distance, and the gradient magnitude is a non-negative real number. This represents the row index of the two-dimensional incremental direction field matrix. The range of values is An integer representing the total number of rows in the two-dimensional incremental direction field matrix; Represents the column index of the two-dimensional incremental direction field matrix. The range of values is The integer representing the total number of columns in the two-dimensional incremental direction field matrix; Indicates the position of the two-dimensional incremental direction field matrix located at the th Line number The horizontal gradient of the matrix elements in the column, where the value of the horizontal gradient is a non-negative real number; Indicates the position of the two-dimensional incremental direction field matrix located at the th Line number The vertical gradient of the matrix elements in the column, where the value of the vertical gradient is a non-negative real number; This represents the square root operation.
[0087] The horizontal gradient is not calculated for the right boundary matrix elements of the two-dimensional incremental directional field matrix, and the vertical gradient is not calculated for the lower boundary matrix elements of the two-dimensional incremental directional field matrix. The matrix elements located in the last row or last column of the two-dimensional incremental directional field matrix do not participate in the subsequent candidate anomaly region labeling.
[0088] In some embodiments, the monitoring area corresponding to the matrix element whose gradient amplitude is greater than a predetermined threshold is extracted. The predetermined threshold is dynamically determined by the weighted sum of the historical mean and standard deviation of the gradient amplitude of each monitoring area under normal boiler operation conditions. Under normal boiler operation conditions, historical gradient amplitude sequences are continuously collected for each matrix element position, and the historical mean and historical standard deviation of the historical gradient amplitude sequence are calculated. The predetermined threshold is set as the sum of the historical mean and the weighting coefficient multiplied by the historical standard deviation. The weighting coefficient is set to 3, based on the fact that the historical gradient amplitude under normal boiler operation conditions approximately follows a normal distribution, and the one-sided confidence probability corresponding to the historical mean plus 3 times the historical standard deviation is 99.87%, which can distinguish gradient amplitudes within the normal fluctuation range through the predetermined threshold. The predetermined threshold is recalculated at each sampling time based on the historical gradient amplitude within the most recent historical window, and the duration of the most recent historical window is preset according to the stability of the boiler operation conditions.
[0089] In some embodiments, a labeling matrix with the same size as the two-dimensional incremental direction field matrix is initialized. Each element in the labeling matrix is initially set to an unlabeled state. The two-dimensional incremental direction field matrix is traversed in row-major order, with the row traversal order increasing from the first row to the total number of rows in the matrix, and the column traversal order increasing from the first column to the total number of columns. When a matrix element with a gradient magnitude greater than a predicted threshold and whose corresponding position in the labeling matrix is unlabeled is encountered, a new connected component number is created. The connected component numbers are integers starting from 1 and incrementing by 1 for each newly created connected component.
[0090] Using the matrix element that creates the new connected component number as the seed point, check the matrix elements adjacent to the seed point in the four directions (up, down, left, and right). When the gradient magnitude of the adjacent matrix element is also greater than the pre-judgment threshold and the corresponding position in the labeling matrix is in an unlabeled state, the adjacent matrix element is added to the current connected component number, and the corresponding position in the labeling matrix is updated to a labeled state. Then, using the newly added matrix element as the new seed point, repeat the check of the matrix elements adjacent to the new seed point in the four directions (up, down, left, and right) until all adjacent matrix elements of the connected component number do not satisfy the condition that the gradient magnitude is greater than the pre-judgment threshold or have all been labeled, at which point the growth process of the current connected component ends.
[0091] Count the number of matrix elements contained under each connected component number. Compare the number of matrix elements for all connected component numbers, and select all monitoring areas corresponding to the connected component with the fewest matrix elements as the labeled objects. Mark all monitoring areas covered by the connected component with the fewest matrix elements as candidate anomaly regions. When multiple connected component numbers have the same minimum number of matrix elements, select the connected component with the smallest connected component number as the labeled object. Store the monitoring area numbers of the candidate anomaly regions in the candidate anomaly region set for subsequent time-series correlation tracing.
[0092] In practice, the noise increment direction vector of the candidate anomaly region at the current sampling time is used as the reference direction vector. The set of candidate anomaly regions from the previous sampling time is obtained and stored in the data processing unit. This set includes the monitoring region number marked as a candidate anomaly region at the previous sampling time and its corresponding noise increment direction vector. The noise increment direction vector with the smallest angle to the reference direction vector is searched within the candidate anomaly region set from the previous sampling time. The angle between two direction vectors is calculated by taking the inner product of the two direction vectors, their magnitudes, and applying the inverse cosine function. If the minimum angle is less than a preset angle threshold, the noise increment direction vector of the candidate anomaly region corresponding to the reference direction vector at the current sampling time is associated with the noise increment direction vector with the smallest angle in the candidate anomaly region set from the previous sampling time as a continuous node on the same tracking chain. A preset angle threshold is stored in the data processing unit. The preset angle threshold is set to 15 degrees. The basis for setting the preset angle threshold to 15 degrees is that, under normal boiler operating conditions, the angle change of the noise increment direction vector in the same monitoring area between adjacent sampling times caused by environmental fluctuations does not exceed 15 degrees. An angle change exceeding 15 degrees indicates that the noise increment direction characteristics have changed significantly. When the minimum value of the direction vector angle is greater than or equal to the preset angle threshold, the noise increment direction vector of the candidate abnormal area corresponding to the reference direction vector at the current sampling time is taken as the starting node of the new tracking chain, and the new tracking chain obtains a new tracking chain number.
[0093] In some embodiments, the duration of a successfully associated tracking chain is incremented by one. The duration of the tracking chain is recorded in the tracking chain's status register. A successfully associated tracking chain is one that already existed at the previous sampling time, and its duration is the value at the previous sampling time plus one sampling interval. When a tracking chain is successfully associated, the noise increment direction vector at the current sampling time is appended to the end of the tracking chain's node sequence.
[0094] In practice, the tracking chain is used to generate a symbol sequence of the noise increment direction vector at each sampling moment within the current duration. For each sampling moment stored on the tracking chain, the frequency index positions where the symbol function output is non-zero are extracted. The non-zero symbol function output value includes both positive and negative signs. The symbol function output values at these frequency index positions are sorted in ascending order by frequency index to form the symbol subsequence for the current sampling moment. This subsequence does not contain zero. The symbol subsequences from multiple consecutive sampling moments on the tracking chain are concatenated in chronological order to form the overall symbol sequence of the tracking chain. The chronological order is the continuous chronological order from the sampling moment of the starting node of the tracking chain to the current sampling moment. The length of the overall symbol sequence of the tracking chain is the sum of the lengths of the symbol subsequences at each sampling moment on the tracking chain.
[0095] In practice, the symbol values of adjacent positions in the overall symbol sequence of the tracing chain are traversed. Whenever an adjacent symbol value changes from positive to negative or vice versa, the flip count is incremented by one. Flips do not increment if the adjacent symbol value changes from positive to positive, from negative to negative, from positive to zero, from negative to zero, from zero to positive, or from zero to negative. After traversal, the flip count is divided by the length of the overall symbol sequence minus one to obtain the current symbol flip frequency of the tracing chain. The current symbol flip frequency of the tracing chain is calculated using the following formula:
[0096]
[0097] in, Indicates the tracking chain number. The value range is an integer representing all tracking chain numbers created in the data processing unit; Indicates the current sampling time sequence number. It is a positive integer greater than or equal to 2; Indicates the tracking chain number is The tracking chain at the sampling time sequence number The current sign-flipping frequency, where the value of the current sign-flipping frequency is a real number greater than or equal to 0 and less than or equal to 1; Indicates the tracking chain number is The tracking chain at the sampling time number The flip count value is an integer greater than or equal to 0. The flip count value is obtained by traversing the entire symbol sequence of the tracing chain and counting the adjacent symbol values. Indicates the tracking chain number is The tracking chain at the sampling time number The length of the overall symbol sequence, which is a positive integer greater than or equal to 2, is the sum of the lengths of the symbol subsequences at each sampling time on the tracking chain; This represents the value after subtracting one from the length of the entire symbol sequence. The value of is a positive integer greater than or equal to 1. When the length of the overall symbol sequence is less than 2, the current symbol flip frequency is not calculated, the current symbol flip frequency is set to 0 and temporarily stored in the status register of the tracking chain.
[0098] The current symbol flip frequency is temporarily stored in the tracking chain's status register, serving as a historical reference value for comparison with the symbol flip frequency at the next sampling time. After data processing is completed for each sampling time, the symbol flip frequency is recalculated based on the updated tracking chain node sequence. The flip frequency attenuation ratio is calculated by combining the regional average flip frequency of the current sampling time with that of the previous sampling time. These two parameters are updated sequentially with the sampling time. A complete numerical solution is performed in each sampling cycle. The calculated parameters are directly compared with a preset lower limit for the flip frequency and a preset attenuation threshold. The entire comparison process relies directly on the parameters output at each sampling time, without using a pre-fixed single value to directly determine leakage risk. The tracking chain's status register stores the tracking chain's duration, node sequence, overall symbol sequence, current symbol flip frequency, and historical symbol flip frequency sequence.
[0099] In practice, the current symbol flip frequency is compared with a preset lower limit. This preset lower limit is stored in the data processing unit and is set to 0.15. The basis for this setting is that, in the historical records of the symbol flip frequency of the noise increment direction vector under normal boiler operating conditions, the symbol flip frequency during normal operation is higher than 0.2. When the noise increment direction characteristics caused by early leakage tend to stabilize, the symbol flip frequency drops below 0.1. The preset lower limit of 0.15 is used to distinguish between normal fluctuations and the decrease in symbol flip frequency caused by early leakage. When the current symbol flip frequency is lower than the preset lower limit, the arithmetic mean of the symbol flip frequencies of all tracking chains within the candidate anomaly region at the current sampling time is calculated, and this arithmetic mean is used as the regional average flip frequency of the candidate anomaly region. The symbol flip frequency of all tracking chains within the candidate anomaly region is read from the status register of each tracking chain at the current sampling time. The arithmetic mean is calculated by summing the current symbol flip frequencies of all tracking chains within the candidate anomaly region and dividing by the number of tracking chains within the candidate anomaly region.
[0100] In some embodiments, the average flip frequency of the candidate anomaly region at the previous sampling time is used as a reference value, and the average flip frequency of the region at the current sampling time is divided by the reference value to obtain the flip frequency attenuation ratio of the candidate anomaly region. The calculation of the flip frequency attenuation ratio is performed when the average flip frequency of the candidate anomaly region at the previous sampling time is greater than 0, and when the reference value is equal to 0, the flip frequency attenuation ratio is set to 0.
[0101] In practical implementation, when the attenuation ratio of the flip-over frequency is less than a preset attenuation threshold, the region identifier of the candidate abnormal region and the corresponding sensor array index table are output as the determination result of the early leakage risk region. The preset attenuation threshold is stored in the data processing unit and is set to 0.6. The basis for setting the preset attenuation threshold to 0.6 is that during the formation of early leakage, the average flip-over frequency of the region shows a significant decrease between adjacent sampling times. When the decrease exceeds 40%, the attenuation ratio of the flip-over frequency is less than 0.6, which can distinguish the slow changes caused by boiler load adjustment. The region identifier of the candidate abnormal region is the monitoring region number corresponding to the candidate abnormal region, and the sensor array index table is the sensor array index table corresponding to the candidate abnormal region. The region identifier and the sensor array index table are read from the data processing unit and pushed to the boiler operation monitoring terminal.
[0102] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.
Claims
1. A method for predicting early leakage of boiler tubes based on regional noise increment feature identification, characterized in that, include: Multiple acoustic sensors deployed inside the boiler furnace are divided into regional grids according to spatial coordinates to obtain sensor arrays for each monitoring area; The time-domain noise signals of the acoustic sensors in each monitoring area are collected synchronously, and a fast Fourier transform is performed on each time-domain noise signal to obtain the frequency-domain noise spectrum of each monitoring area. Within each monitoring area, the current frequency domain noise spectrum is compared with the frequency domain noise spectrum at the previous sampling time by frequency point differential calculation to obtain the differential amplitude at each frequency point, and the positive and negative sign sequence of the differential amplitude at each frequency point is counted to generate the noise increment direction vector of the area. An incremental direction field matrix is constructed using the noise increment direction vectors of all monitoring areas. Spatial gradient operation is performed on the incremental direction field matrix to obtain the connected domain of the region boundary where the direction gradient magnitude exceeds the pre-judgment threshold. The monitoring area covered by this connected domain is marked as a candidate anomaly region. Within the candidate anomaly region, the noise increment direction vectors at multiple consecutive sampling times are time-series correlated and tracked. When the sign flip frequency of the tracked noise increment direction vector is lower than a dynamic threshold, the dynamic threshold is a composite judgment threshold, which includes a preset lower limit of the flip frequency and a preset attenuation threshold. When the sign flip frequency of the tracked noise increment direction vector is lower than the preset lower limit of the flip frequency, and the flip frequency attenuation ratio corresponding to the candidate anomaly region is less than the preset attenuation threshold, the candidate anomaly region is determined to be an early leakage risk region.
2. The method for predicting early leakage of boiler tubes based on regional noise increment feature identification according to claim 1, characterized in that, Multiple acoustic sensors deployed inside the boiler furnace are divided into regional grids according to spatial coordinates to obtain sensor arrays for each monitoring area, specifically including: Extract the three-dimensional spatial installation coordinates of all acoustic sensors in the boiler furnace, and divide the acoustic sensors into several sensor groups with the constraint that the Euclidean distance between adjacent sensors does not exceed a preset spacing threshold. Take the smallest circumscribed cuboid of the three-dimensional spatial installation coordinates of all acoustic sensors in each sensor group, divide the smallest circumscribed cuboid into equal-distance grids in the horizontal and vertical directions to obtain several grid cells, and merge the grid cells containing at least one sensor into the same monitoring area. The sensor numbers of the acoustic sensors contained in each monitoring area are associated and stored to form a sensor array index table corresponding to each monitoring area.
3. The method for predicting early leakage of boiler tubes based on regional noise increment feature identification according to claim 1, characterized in that, The frequency domain noise spectrum of each monitoring area is obtained, specifically including: According to the sensor array index table, the analog-to-digital converters of all acoustic sensors in each monitoring area are synchronously triggered with the same sampling clock source, and the discrete time-domain sampling sequence within a preset time period is collected as the time-domain noise signal of the sensor. The time-domain noise signals of each acoustic sensor are truncated by applying a Hanning window function. The truncated time-domain sequences are then padded with zeros to the preset number of Fourier transform points and a fast Fourier transform is performed to obtain the initial frequency domain sequences of each sensor. The frequency domain amplitude of the monitoring area at the same frequency point is obtained by taking the arithmetic mean of the amplitudes of the initial frequency domain sequences of all sensors in the same monitoring area at the same frequency point. After aggregating all frequency points, the frequency domain noise spectrum of each monitoring area is formed.
4. The method for predicting early leakage of boiler tubes based on regional noise increment feature identification according to claim 1, characterized in that, Generating the noise increment direction vector for this region specifically includes: For the same monitoring area, extract the frequency domain noise spectrum vector at the current sampling time and the frequency domain noise spectrum vector at the previous sampling time, subtract the frequency domain amplitude at the same frequency index in the two vectors, and obtain the differential amplitude corresponding to each frequency index. For each frequency index, a sign function is used to calculate the differential amplitude. When the differential amplitude is greater than zero, the sign function outputs a positive sign; when the differential amplitude is less than zero, the sign function outputs a negative sign; and when the differential amplitude is equal to zero, the sign function outputs zero. The sign function outputs of all frequency indices are arranged in ascending order of frequency index to form the noise increment direction vector of the monitoring area.
5. The method for predicting early leakage of boiler tubes based on regional noise increment feature identification according to claim 1, characterized in that, The incremental direction field matrix is constructed using the noise increment direction vectors of the entire monitoring area, specifically including: The noise increment direction vectors of each monitoring area are used as matrix elements and arranged into a two-dimensional increment direction field matrix according to the spatial adjacency topology of each monitoring area. The row and column indices of this matrix correspond to the sequence numbers of the monitoring areas in the horizontal and vertical directions, respectively.
6. The method for predicting early leakage of boiler tubes based on regional noise increment feature identification according to claim 5, characterized in that, Performing spatial gradient calculations on the incremental direction field matrix yields the connected regions whose directional gradient magnitudes exceed a predetermined threshold, specifically including: For each element in the two-dimensional incremental direction field matrix, the Hausdorff distance between its direction vector and the direction vector of its right-hand neighbor and its lower neighbor is taken as the horizontal gradient and vertical gradient, respectively, and the gradient magnitude of the matrix element is synthesized from the horizontal gradient and the vertical gradient. The monitoring areas corresponding to matrix elements whose gradient magnitude is greater than the predicted threshold are extracted. The four-connected neighborhood growth algorithm is used to merge adjacent extracted monitoring areas into connected regions. All monitoring areas covered by the connected region with the smallest area are marked as candidate abnormal regions.
7. The method for predicting early leakage of boiler tubes based on regional noise increment feature identification according to claim 6, characterized in that, A four-connected neighborhood growing algorithm is used to merge adjacent extracted monitoring regions into connected components. The entire monitoring region covered by the connected component with the smallest area is marked as a candidate anomaly region, specifically including: Initialize a labeling matrix with the same size as the two-dimensional incremental direction field matrix. The initial value of each element in the labeling matrix is unlabeled. Traverse the two-dimensional incremental direction field matrix in row-first and column-second order. When the gradient magnitude is greater than the prediction threshold and the corresponding position in the labeling matrix is unlabeled, create a new connected component number. Using the matrix element as the seed point for growth, check the matrix elements adjacent to the seed point in the four directions of up, down, left, and right. When the gradient magnitude of the adjacent matrix element is also greater than the pre-judgment threshold, the adjacent matrix element is assigned to the current connected component number and the corresponding position in the label matrix is updated to the labeled state. Then, the newly assigned matrix element is used as the new seed point to repeat the growth process. Count the number of matrix elements contained under each connected component number, take all the monitoring areas corresponding to the connected component with the fewest matrix elements as the marking objects, and mark these monitoring areas as candidate abnormal areas.
8. The method for predicting early leakage of boiler tubes based on regional noise increment feature identification according to claim 1, characterized in that, Within the candidate anomaly region, temporal correlation tracking is performed on the noise increment direction vectors at multiple consecutive sampling times, specifically including: Using the noise increment direction vector of the candidate anomaly region at the current sampling time as a reference, search for the noise increment direction vector with the smallest angle with the reference direction vector in the candidate anomaly region set at the previous sampling time. If the smallest angle with the reference direction vector is less than a preset angle threshold, then the two direction vectors are associated as consecutive nodes on the same tracking chain. Increment the duration of the successfully associated tracking chain by one, and count the symbol sequence of the noise increment direction vector at each sampling moment within the current duration of the tracking chain. Calculate the ratio of the number of times adjacent symbols in the symbol sequence flips to the total length of the symbol sequence to obtain the current symbol flip frequency of the tracking chain.
9. The method for predicting early leakage of boiler tubes based on regional noise increment feature identification according to claim 8, characterized in that, To obtain the current symbol flip frequency of the tracing chain, specifically including: For each sampling moment corresponding to the noise increment direction vector stored on the tracking chain, extract the frequency index positions where the symbol function output is non-zero in each vector, take the symbol function output values at these frequency index positions and arrange them in ascending order of frequency index to form the symbol subsequence of that sampling moment, and splice the symbol subsequences of multiple consecutive sampling moments on the tracking chain in chronological order to form the overall symbol sequence of the tracking chain; Traverse the symbol values of adjacent positions in the overall symbol sequence. Whenever an adjacent symbol value changes from positive to negative or vice versa, increment the flip count by one. After traversing the sequence, divide the flip count by the length of the overall symbol sequence and decrement by one to obtain the current symbol flip frequency of the tracing chain. The current symbol flip frequency is temporarily stored in the status register of the tracking chain, serving as a historical reference value for comparison of the symbol flip frequency at the next sampling moment.
10. The method for predicting early leakage of boiler tubes based on regional noise increment feature identification according to claim 8, characterized in that, When the sign-flipping frequency of the tracked noise increment direction vector is lower than the dynamic threshold, the candidate anomaly region is determined to be an early leakage risk region, specifically including: Compare the current symbol flip frequency with the preset lower limit of the flip frequency. When the current symbol flip frequency is lower than the preset lower limit of the flip frequency, the candidate abnormal area is identified as an early leakage risk area. The dynamic threshold, which includes the symbol flip frequency and the flip frequency attenuation ratio used in the determination, are time-series variables calculated in real time at each sampling moment, unlike a fixed static threshold.
11. The method for predicting early leakage of boiler tubes based on regional noise increment feature identification according to claim 10, characterized in that, The current symbol flip frequency is compared with a preset lower limit. When the current symbol flip frequency is lower than the preset lower limit, the candidate anomaly region is identified as an early leakage risk region, specifically including: Calculate the arithmetic mean of the symbol flip frequencies of all tracking chains within the candidate anomaly region at the current sampling time, and use this arithmetic mean as the regional average flip frequency of the candidate anomaly region. Using the average flip frequency of the candidate anomaly region at the previous sampling time as the benchmark value, the average flip frequency of the region at the current sampling time is divided by the benchmark value to obtain the flip frequency attenuation ratio of the candidate anomaly region. When the attenuation ratio of the flip frequency is less than the preset attenuation threshold, the region identifier of the candidate abnormal region and the corresponding sensor array index table are output as the determination result of the early leakage risk region.
12. The method for predicting early leakage of boiler tubes based on regional noise increment feature identification according to claim 6, characterized in that, The predicted threshold is dynamically determined by the weighted sum of the historical mean and standard deviation of the gradient amplitude of each monitoring area under normal boiler operating conditions.
13. The method for predicting early leakage of boiler tubes based on regional noise increment feature identification according to claim 6, characterized in that, The Hausdorff distance is used to measure the maximum difference between two direction vectors.
14. The method for predicting early leakage of boiler tubes based on regional noise increment feature identification according to claim 7, characterized in that, In the four-connected neighborhood growth algorithm, growth in that direction stops when a boundary or a marked element is encountered during the growth process.
15. The method for predicting early leakage of boiler tubes based on regional noise increment feature identification according to claim 8, characterized in that, The preset angle threshold ranges from 0° to 90° and is used to limit the upper limit of the allowable deviation of the angle between direction vectors in time-series correlation tracking.
